Addressing domain shift via imbalance-aware domain adaptation in embryo development assessment

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning models in medical imaging struggle with domain shift and class imbalance.
• Domain shift leads to poor performance in new clinical settings.
• Class imbalance results in underrepresentation of certain disease conditions.

Purpose of the Study:

• To present imbalance-aware domain adaptation (IADA), a novel framework addressing both domain shift and class imbalance.
• To develop a robust AI system for medical imaging applicable in diverse clinical settings.

Main Methods:

• IADA utilizes adaptive feature learning with class-specific attention mechanisms.
• It employs balanced domain alignment with dynamic weighting.
• Adaptive threshold optimization is also incorporated.

Main Results:

• IADA achieved up to 25.19% higher accuracy in embryo development assessment across four imaging modalities.
• The framework maintained balanced performance across classes.
• Robust generalization was observed in low-quality imaging scenarios, with AUC improvements up to 12.56%.

Conclusions:

• IADA effectively tackles domain shift and class imbalance in medical imaging.
• The framework shows potential for developing reliable and equitable medical imaging systems.
• IADA offers improved accuracy and generalization for AI in diverse clinical environments.